Re:LLMs can't self-correct in reasoning tasks, DeepMind stud...
by fragmede on Monday October 09, 2023 @04:55PM
(#37825383)
ChatGPT will happily tell you it doesn't know anything after January 2022, and it'll say "jdnrirmd-gurlksjd" is not a recognizable term, instead of trying to bullshit a definition of that, if that's your bar for "understanding",
Re:LLMs can't self-correct in reasoning tasks, DeepMind stud...
by jiggawatts on Monday October 09, 2023 @05:36PM
(#37825834)
Current LLMs use an optimisation of processing data one word at a time. This makes them bad at solving puzzles involving individual letters or digits.
This is a well known, well understood limitation that can be overcome (Facebook punished a hierarchical model that can parse individual characters), but this technique isn’t used by ChatGPT.
Whenever I see a criticism like this, it says more to me about the hubris of humans and the falsify of their assumed superior intelligence.
You’re criticising the intelligence of a thing you don’t understand yourself!
You didn’t “read the manual”. You didn’t go find out why the AI is failing. You just spouted an angry comment and gave up without gaining any understanding.
PS: the “stupid AI” understands all of this. Just ask it to explain it to you: https://chat.openai.com/share/7bc1c1cb-d888-4ccb-904a-79e9ae...
There were about 80 possible. It maxed out at 19 or so.
Now clearly this is a hard task. Humans can’t do it well and none of the text corpus would offer training on this sort of thing.
If I asked a human to do it they would say “here’s what I got, I don’t think that is all of them”
ChatGPT would apologize when corrected and make a new, equally wrong list with full confidence.
That’s how I interpret understanding. Self awareness and doubt.
Oddly enough the code interpreter seems much stronger on this point. Now that I think about it I don’t think I was using it for the letter combo task.